Fix ValidationError on early client disconnect and ZeroDivisionError in draft metrics logging

[nakaken3013-code]

Two small, independent robustness fixes for edge cases we hit while load-testing the exllamav3 backend under concurrency. Each is a one-line change; happy to split into two PRs if you'd prefer.

---

1. ValidationError when a client disconnects during prefill

Problem
When a streaming chat-completion client disconnects before the first token is generated (i.e. the request is still in prefill), the server raises a pydantic ValidationError instead of ending the request cleanly.

In _chat_stream_collector (endpoints/OAI/utils/chat_completion.py), generation is initialized to {} and the "index" key is only assigned inside the async for generation in new_generation: loop. If the disconnect happens during prefill, the generator yields nothing, the loop body never runs, and the empty {} is returned. _compose_response then constructs ChatCompletionRespChoice(index=None), which fails validation (index must be an int):

ValidationError: index
  Input should be a valid integer [type=int_type, input_value=None]

Fix
Initialize generation = {"index": task_idx} so a valid index is always present even when zero tokens are produced. The in-loop assignment is unchanged.

Repro
Start a streaming chat completion with a large prompt, then abort the client (close the tab / Ctrl-C curl / Open WebUI stop button) while the server is still prefilling, before any token streams back. Before: traceback above. After: the request ends cleanly.

---

2. ZeroDivisionError in draft accept-rate logging

Problem
log_metrics (common/gen_logging.py) computes accept_rate = accept / total_draft with total_draft = draft_accept + draft_reject. When both are 0 (a request that produced no draft tokens while speculative / MTP decoding is enabled), total_draft == 0 and the division raises ZeroDivisionError, aborting the metrics log line.

Fix

accept_rate = accept / total_draft if total_draft > 0 else 0.0

Reports 0.0% in that case instead of crashing.

---

Both were observed with the exllamav3 backend, TP + MTP, under ~8-way concurrent load with clients that cancel mid-request.